Meta is reshaping its AI infrastructure strategy through a sweeping hardware partnership that centers on nvidia cpus alongside next-generation GPUs.
Summary
Meta signs multiyear Nvidia deal spanning GPUs and standalone CPUs
The Facebook parent company Meta has signed a multiyear agreement with Nvidia to buy millions of chips, covering both GPUs and, for the first time, standalone CPUs. The deal includes current Blackwell GPUs, upcoming Rubin GPUs, and the new Grace and Vera processors as stand-alone products. However, neither side has disclosed the total value of the contract.
Ben Bajarin, CEO and principal analyst at tech consultancy Creative Strategies, estimated the package would be worth billions of dollars. Moreover, technology outlet The Register reported that the agreement is likely to add tens of billions to Nvidia’s bottom line over its term. This underscores how aggressively Meta is scaling its AI footprint.
Meta CEO Mark Zuckerberg had already flagged this shift in spending priorities. He announced that Meta plans to almost double its AI infrastructure investment in 2026, with total outlays potentially reaching $135 billion. That said, the new chip deal gives the market a clearer picture of where much of that capital will go.
Nvidia CPU strategy pivots toward inference workloads
The most striking element of the agreement is not the GPU purchase but Meta’s decision to adopt Nvidia’s CPUs at large scale as standalone products. Until early 2026, the Grace processor was offered almost exclusively as part of so-called Superchips, which combine a CPU and GPU on a single module. However, Nvidia officially changed its sales strategy in January 2026 and started selling these CPUs separately.
The first publicly named standalone CPU customer at that time was neocloud provider CoreWeave. Now Meta is joining that list, signaling growing demand for flexible CPU-based architectures. This aligns with a broader transition in AI from training massive models to serving them in production environments.
The company is targeting the fast-expanding inference segment. In recent years, the AI sector focused heavily on GPU-intensive training of large models. However, the emphasis is increasingly shifting to inference, the process of running and scaling those trained systems. For many inference tasks, traditional GPUs are overkill in terms of cost and power.
“We were in the ‘training’ era, and now we are moving more to the ‘inference era,’ which demands a completely different approach,” Bajarin told the Financial Times. That said, this shift does not eliminate GPU demand; instead, it changes the balance between GPU vs CPU workloads inside hyperscale data centers.
Grace and Vera CPUs: technical details and Meta’s deployment plans
Ian Buck, Nvidia’s VP and General Manager of Hyperscale and HPC, said, according to The Register, that the Grace processor can “deliver 2x the performance per watt on those back end workloads” such as running databases. Moreover, he noted that “Meta has already had a chance to get on Vera and run some of those workloads, and the results look very promising.” This highlights Nvidia’s push to optimize power efficiency for large-scale inference and data processing.
The Grace CPU features 72 Arm Neoverse V2 cores and uses LPDDR5x memory, which provides advantages in bandwidth and space-constrained environments. By contrast, Nvidia’s next-generation Vera CPU brings 88 custom Arm cores with simultaneous multi-threading and built-in confidential computing capabilities. These specifications underline Nvidia’s ambition to compete directly with entrenched server CPU vendors.
According to Nvidia, Meta plans to use Vera for private processing and AI features in its WhatsApp encrypted messaging service. Vera deployment is planned for 2027, indicating a multi-year roadmap for Meta’s back-end modernization. However, the company has not provided detailed rollout timelines for each data center region or specific services beyond messaging and security-related workloads.
Competitive landscape: Nvidia enters the server CPU arena
Nvidia’s move to sell CPUs as standalone products puts it in direct competition with Intel and AMD in the lucrative server market. Previously, much of Nvidia’s growth came from GPUs, but the addition of CPUs gives the company a more complete data center portfolio. Moreover, it allows customers to build full stacks around the same vendor rather than mixing components from multiple suppliers.
By purchasing standalone Nvidia CPUs, Meta is deviating from the strategy pursued by other hyperscalers. Amazon relies on its own Graviton processors, while Google leans on its custom Axion chips. Meta, by contrast, is buying from Nvidia even as it continues to design its own AI accelerators. However, the Financial Times reported that Meta’s internal chip efforts have “suffered some technical challenges and rollout delays.”
For Nvidia, the competitive pressure is also intensifying. Google, Amazon, and Microsoft have each announced new in-house chips over recent months. In parallel, OpenAI has co-developed a processor with Broadcom and signed a significant supply agreement with AMD. Several startups, including Cerebras, are pushing specialized inference silicon that could erode Nvidia’s dominance if widely adopted.
Market tensions, stock reactions, and multi-vendor strategies
In December, Nvidia acquired talent from inference chip company Groq in a licensing deal, aiming to reinforce its technology base in this new inference era computing phase. However, investor sentiment remains sensitive to any sign of customer diversification. Late last year, Nvidia’s stock fell four percent after reports suggested Meta was in talks with Google about using Tensor Processing Units. No formal agreement on TPUs has been announced since.
Meta is also not locked exclusively into Nvidia hardware. According to The Register, the company operates a fleet of AMD Instinct GPUs and took part directly in designing AMD’s Helios rack systems, which are scheduled for release later this year. Moreover, this multi-vendor approach gives Meta leverage in price negotiations and helps reduce supply risk across its fast-growing meta ai infrastructure.
As the company expands its data centers, the question “does nvidia sell cpus” is being answered in practice through deployments like this one. The broader meta nvidia agreement showcases how nvidia cpus are becoming a central piece of large-scale inference architectures, even as hyperscalers experiment with their own custom silicon and rival accelerator platforms.
In summary, Meta’s multiyear hardware deal underscores a structural transition in AI from training-heavy GPU clusters to inference-optimized architectures built around advanced CPUs such as Grace and Vera. However, with Intel, AMD, cloud-native processors, and specialized startups all competing for the same workloads, Nvidia faces a complex battle to turn its new CPU strategy into long-term data center dominance.
